Dan
32 posts
Dan
@dan_8bit
Mechanical/aerospace engineer. Building AI agents for code-first engineering. Quote-ready CAD, durable workflows, and traceable run records.
Katılım Haziran 2026
51 Takip Edilen12 Takipçiler
Looks like this was just testing done by OpenAI as they work out optimal usage, investigate limit drains, and so on.
For me, GPT 5.6 (on multiple different thinking/reasoning levels) has been working just fine throughout the day... with no noticeable issues or any degraded output. So all is good on that front.
Dan@dan_8bit
Hmm... Maybe this is why Sol has felt different today. I was just posting about how medium thinking level was "finally feeling like it was working... Not spinning it's wheels" Could it be the effort level changes? 🤔
English
Hmm... Maybe this is why Sol has felt different today.
I was just posting about how medium thinking level was "finally feeling like it was working... Not spinning it's wheels"
Could it be the effort level changes? 🤔
Lentils@Lentils80
🚨 GPT-5.6 Sol's juice values (thinking budgets) have been severely degraded compared to release day If Sol now feels faster and more "efficient", this is probably why Terra and Luna juice values aren't affected, so their thinking budgets are now higher than Sol's
English
Hmm.. I still think there might be a bug in how they calculaate cached/uncached tokens for Grok 4.5 right now, not 100% sure though.
But for context... to hit my weekly limit, that was about ~$150 of API equivalent usage on uncached tokens for me.
So in 4 weeks that would allow ~$600 in the subscription. Thing is, you would have to wait for a reset if you use it too fast.. so it's a tough decision.
English
@usr_bin_roygbiv Cerebras would be:
- Mean TTFT < 1 second
- Mean tok/s > 700
- A duration of multiple seconds+
English
@usr_bin_roygbiv I understand this perfectly well.. and have seen/read when it's expected...
But this right here is an Omp Bench issue, not yet Cerebras.
English
Comments on this are making me realize why people still think anth has a chance in hell.
Look up what cerebras does
Roy@usr_bin_roygbiv
so uhhhh
English
Right, what I'm saying is, it was "thinking" in the background for ~7 - 12 seconds, then bursted a stream of the full output in the final 1-2 secs.
This makes the tok/sec measured by omp bench misleadingly high.
The actual normalized tok/sec when going with those outputs of 1-2k, is around 113 - 231 tok/sec.
Still faster than default Codex, but not quite 750.
English
@SkinnyfatTony I'm coding an application to build mechanical stuff for me - where does that fall on the list? 😅
English
I'm enjoying vibe coding useful applications more than building mechanical stuff. Lawd help me.
English
@usr_bin_roygbiv Naw I take this back, doesn't make sense.
When you divide the output by TTFT, it just gives the same speed/rate as the original.
It looks like it just streams all the tokens at once. Which distorts the calculation.
@usr_bin_roygbiv
English
@TheHauskarl It might be using Grok Build directly vs the API in another harness, as someone here just mentioned.
Seems the limits are more generous through Build than the API.
English
There's something not right there. I've been watching usage carefully since the weekly limits came in to better keep a grasp on what I've got left.
An extremely long, deep Grok Build session interacting directly with my PC over multiple hours I think came to less than 10%.
Have you tried compacting your chats? It saves tokens. Asking multiple questions in one call has the same effect. Several other tricks to maximise efficiency and I haven't made full use of them even.
The only part of weekly limits I'll agree are absurd is Grok Imagine. You could comfortably max out your weekly limit in an hour if you just spam a few videos / images.
English
@scholar1013 Yes, this does appear to be the case... My usage here was through API in a different harness.
English
@dan_8bit i now know the reason why people consumes tokens so differently. if you use subscription in grok build (grok's own tui), there are way more tokens in the plan. If you use other harness, the usage is labeled as API (rather than grok build), and this has lower token budget
English
Based on my token usage for Grok 4.5, this essentially means:
A SuperGrok Heavy account currently allows 350M Grok 4.5 tokens in its weekly limit? 🤔
I may try Cursor next.
Dan@dan_8bit
Switched over to Grok 4.5 for a few days for some heavy coding usage... completely hit the weekly limit in 2 days! It's a great model, but the SuperGrok Heavy Limit seems very low for the price... even Codex has far more usage.
English
@brennen_harris_ Very large codebase, being worked on by several subagents at one time.
Do you know how many tokens you used?
If my weekly SuperGrok Heavy estimate is accurate at ~350M tokens, then 20% of a sub is about 70M total tokens of usage.
English
Genuinely how is this possible?
I had 4 days this week of intense coding usage on my Grok Heavy sub in Grok Build and used less than 20% of my sub. I’m not saying it isn’t happening to people out there. But either there’s a massive difference in token efficiency between users, or they are measuring volumes differently for different accounts.
English
@scholar1013 Yes, agreed. I'm currently trying Cursor for Grok 4.5, and it is already clear this is going to allow significantly more tokens for the price...
English
@dan_8bit oh, that's way more below my expectation, even chatgpt pro lite has more weekly tokens
English
@scholar1013 Actually this 358M includes uncached.
So this means... a SuperGrok Heavy account currently allows 350M Grok 4.5 tokens? 🤔
In the weekly limit, that is.
English
@scholar1013 2 days of Grok 4.5:
- 7.8M uncached tokens
- 358M cached tokens
English